DID-M3D: Decoupling Instance Depth for Monocular 3D Object Detection

Monocular 3D detection has drawn much attention from the community due to its low cost and setup simplicity. It takes an RGB image as input predicts boxes in space. The most challenging sub-task lies instance depth estimation. Previous works usually use a direct estimation method. However, this paper we point out that on is non-intuitive. coupled by visual clues attribute clues, making it hard be directly learned network. Therefore, propose reformulate combination of surface (visual depth) (attribute depth). related objects’ appearances positions image. By contrast, relies inherent attributes, which are invariant object affine transformation Correspondingly, decouple location uncertainty into uncertainty. combining different types depths associated uncertainties, can obtain final depth. Furthermore, data augmentation monocular limited physical nature, hindering boost performance. Based proposed disentanglement strategy, alleviate problem. Evaluated KITTI, our method achieves new state-of-the-art results, extensive ablation studies validate effectiveness each component codes released at https://github.com/SPengLiang/DID-M3D .